TWON: Ranker - Modularized & Weighted Timeline Ranking

todo

Project Setup

todo

# install Python requirements
make install

# start api with hot-reload for development
make dev

# start api for production
make serve

# run unit tests
make test

Modules

The current TWON ranking/recommendation algorithm is divided into three encapsulated modules that composed denote the ranking function. These modules provide the functionalities to measure engagement given a predefined data format, represent the course of time, and partly randomize the final ranking.

Noise

We draw random floating point numbers from the normal distribution provided lower and upper boundaries to generate a multiplicative noise (the neutral value defined as LOW = HIGH = 1. will result in no noise).

from src.modules import Noise

LOW: float
HIGH: float
N: int

eps = Noise(low=LOW, high=HIGH)

rnd_number: float = eps()
rnd_samples: List[float] = eps.draw_samples(N)

Decay

We compute a decay factor based on the time elapsed between two references. In the context of this project, the decay factor decreases the relevance of posts over time. That results in older posts without recent interaction being less often recommended. We instantiate a decay object by defining a minimum value that serves as a lower boundary for the decay and a reference time interval. When called, the decay object calculates a time difference between the reference time interval and the observed time interval. The maximum computed value is defined as 1, for observation == reference.

from src.modules import Decay

MINIMUM: float
REFERENCE_TIMEDELTA: timedelta

decay = Decay(minimum=MINIMUM, reference_timedelta=REFERENCE_TIMEDELTA)

decay_factor: float = decay(observation_datetime=datetime, reference_datetime=datetime) 

Engagement

The engagement module computes a score based on a plain count of observations count_based or the sum of decayed values for the individual data points decay_based. For the decayed-based version, an instantiated decay module is necessary. Optionally, the output can be normalized with the natural logarithm log_normalize.

from src.modules import Engagement

FUNC: Literal['count_based', 'decay_based']
LOG_NORMALIZE: bool

E = Engagement(func=FUNC, log_normalize=LOG_NORMALIZE)

score_count: int = E(items=List[datetime])
score_decay: float = E(items=List[datetime], reference_datetime=datetime, decay=decay)

Usage

todo

Post

We model a social media post for the TWON simulation with the following class. The object contains the following attributes:

• id: A unique identifier (ID) of the post as a string.
• timestamp: The timestamp containing the post creation date and time. The class expect a string formatted defined by ISO 8601.
• likes/dislikes: A list of observations denoted as timestamps (see above).
• comments: A list of Post objects representing comments. This approach allows arbitrary nest posts into complex tree structures for future TWON modifications. The current implementation ignores those sublevel structures and only counts direct comments of the main posts into the observations.

from src.post import Post

ID: str
TIMESTAMP: datetime
OBERSERVATIONS: List[datetime]
COMMENTS: List[Post] # post objects w/o comments

post = Post(
    id=ID,
    timestamp=TIMESTAMP,
    likes=OBERSERVATIONS,
    dislikes=OBERSERVATIONS,
    comments=COMMENTS,
)

Request

The Request object denotes the attributes needed for interaction with the ranker. It collates all modules previously defined and combines them with weights for each observation type. The object contains the following attributes:

• items: A list of Post objects
• reference_datetime: The reference timestamp (defaults to now while receiving the request).
• decay: Attributes needed to instantiate the Decay module
• noise: Attributes needed to instantiate the Noise module
• engagement: Attributes needed to instantiate the Engagement module
• weights: Weights to tweak the different engagement factors (defaults 1.0 for all)

from src.request import Request, Weights

ITEMS: List[Post] # see Usage:Post
REFERENCE_DATETIME: datetime
DECAY: modules.Decay  # see Modules:Decay
NOISE: modules.Noise  # see Modules:Noise
ENGAGEMENT: modules.Engagement # see Modules:Engagement

WEIGHTS: Weights(
    likes=1.
    dislikes=1.
    comments=1.

    comments_likes=1.
    comments_dislikes=1.
)

req = Request(
    items=ITEMS,
    reference_datetime=REFERENCE_DATETIME,
    decay=DECAY,
    noise=NOISE,
    engagement=ENGAGEMENT
    weights=WEIGHTS
)

Ranker

todo

# todo